What is Nautir
Nautir is DevZero's LLM inference platform -- a gateway in your request path plus a control plane that governs, measures and evaluates what flows through it.
What is Nautir
Nautir is DevZero's inference platform. You point your LLM traffic at a gateway, and the platform gives you spend attribution, guardrails, routing control and model evaluation over that traffic -- with no change to your application beyond a base URL and a key.
The three components
| Component | What it does | Where it runs |
|---|---|---|
| AI Gateway | Sits between your clients and the LLM providers. Applies governance, caching, routing and optimization, then forwards the request. | Hosted by DevZero, or self-hosted in your own environment |
| Control plane | Holds guardrails, routing profiles, keys, credentials, the model catalog, trials, datasets and telemetry. Serves the dashboard. | DevZero |
| Dashboard | The Inference Platform section of the web application, where you configure the above and read the results. | DevZero |
The control plane is never in the request path. It answers the gateway's questions -- may this request spend, which profile applies, which trials are armed -- and the gateway caches those answers so a control-plane outage degrades rather than denies.
What the gateway does to a request
A request arriving at the gateway is classified by the credential it carries, and the classification decides everything that follows.
| Credential shape | Treated as | Behaviour |
|---|---|---|
A DevZero-issued key (sk-dz-...) | Keyed product traffic | Governed, routed, optimized, metered. Reaches the product surfaces. |
A DevZero access token with the inference scope | Keyed product traffic | Same as above; the gateway verifies the token as a resource server of DevZero's authorization server. |
| Your own provider credential, or a subscription bearer (a Claude seat token, a ChatGPT JWT) | Passthrough traffic | Forwarded byte for byte in both directions, on your own credential, with two deliberate compatibility exceptions on the Anthropic-native surface. No DevZero header is added. |
That split is permanent and deliberate. It is what lets a Claude Code seat and an API-key integration share one gateway deployment -- even the same listener, method and path -- without either behaviour leaking into the other.
The passthrough listeners are not part of the AI Gateway product surface. The product surface is an explicit allowlist of routes; everything else on the gateway keeps its existing behaviour and is outside the published contract.
Sources: what the platform is watching
Everything that reports LLM telemetry for your team is a source, and every source is exactly one kind:
- A proxy gateway routes requests through an installed gateway process. It has an installation record, a provisioned URL and a connectivity status, and it reports liveness by pinging home.
- A direct source is an SDK or client reporting telemetry directly, with no proxy process in the request path. It has no installation and no deployment method; its activity is measured from request telemetry rather than from heartbeats.
A proxy gateway records how it was deployed -- hosted, helm, docker or binary -- and reports that itself. Hosted means DevZero runs the process; every other method means you do.
Residency: content uplink
Whether your prompt and response text may leave the environment the gateway runs in is one control, content uplink. It is set by environment configuration only, never from the dashboard, and it defaults from the deployment method: on for hosted, content-free for every self-hosted method.
Billing regimes
Every request is classified into one of two billing regimes by the shape of its credential, and the regime decides how value is denominated.
| Regime | Credential | Value is denominated in |
|---|---|---|
| Metered | Per-token API keys (your own, or DevZero-issued) | Dollars. Compression saves real money and the full evidence loop applies. |
| Subscription | Prepaid seat traffic (OAuth-shaped credentials) | Rate-limit headroom and cache stability. Tokens not spent extend how far a prepaid seat's limits stretch. |
Pricing saved subscription tokens at API rates is allowed only as an estimate, never as a measured claim, because no bill moved. See AI spend for how the product keeps that distinction visible.
What this documentation covers
- Gateway -- installing a gateway, pointing traffic at it, how keys work, and what you see afterwards.
- AI spend -- spend, savings, efficiency and attribution, and the honesty rules attached to each figure.
- Routing -- routing profiles, fallback chains, provider preferences, auto mode and the model catalog.
- Guardrails -- budgets, ceilings, model and provider access, and content rules.
- Evals -- trials, experiments, datasets, compare and watches.